Accumulo fully recovered when I restarted the loggers. Very impressive.
On Mon, Jan 28, 2013 at 9:32 AM, John Vines <[email protected]> wrote: > And make sure the loggers didn't fill up their disk. > > Sent from my phone, please pardon the typos and brevity. > On Jan 28, 2013 8:54 AM, "Eric Newton" <[email protected]> wrote: > >> What version of accumulo was this? >> >> So, you have evidence (such as a message in a log) that the tablet server >> ran out of memory? Can you post that information? >> >> The ingested data should have been captured in the write-ahead log, and >> recovered when the server was restarted. There should never be any data >> loss. >> >> You should be able to ingest like this without a problem. It is a basic >> test. "Hold time" is the mechanism by which ingest is pushed back so that >> the tserver can get the data written to disk. You should not have to >> manually back off. Also, the tserver dynamically changes the point at >> which it flushes data from memory, so you should see less and less hold >> time. >> >> The garbage collector cannot run if the METADATA table is not online, or >> has an inconsistent state. >> >> You are probably seeing a lower number of tablets because not all the >> tablets are online. They are probably offline due to failed recoveries. >> >> If you are running Accumulo 1.4, make sure you have stopped and restarted >> all the loggers on the system. >> >> -Eric >> >> On Mon, Jan 28, 2013 at 8:28 AM, David Medinets <[email protected] >> >wrote: >> >> > I had a plain Java program, single-threaded, that read an HDFS >> > Sequence File with fairly small Sqoop records (probably under 200 >> > bytes each). As each record was read a Mutation was created, then >> > written via Batch Writer to Accumulo. This program was as simple as it >> > gets. Read a record, Write a mutation. The Row Id used YYYYMMDD (a >> > date) so the ingest targeted one tablet. The ingest rate was over 150 >> > million entries for about 19 hours. Everything seemed fine. Over 3.5 >> > Billion entries were written. Then the nodes ran out of memory and >> > Accumulo nodes went dead. 90% of the server was lost. And data poofed >> > out of existence. Only 800M entries are visible now. >> > >> > We restarted the data node processes and the cluster has been running >> > garbage collection for over 2 days. >> > >> > I did not expect this simple approach to cause an issue. From looking >> > at the logs file, I think that at least two compactions were being run >> > while still ingested those 176 million entries per hour. The hold >> > times started rising and eventually the system simply ran out of >> > memory. I have no certainty about this explanation though. >> > >> > My current thinking is to re-initialize Accumulo and find some way to >> > programatically monitoring the hold time. The add a delay to the >> > ingest process whenever the hold time rises over 30 seconds. Does that >> > sound feasible? >> > >> > I know there are other approaches to ingest and I might give up this >> > method and use another. I was trying to get some kind of baseline for >> > analysis reasons with this approach. >> > >>
